Eight machines. Forty original repos. One recursive learning loop.
The dp-web4 research collective builds a trust-native ontology — a shared vocabulary and trust model for AI agents, in the knowledge-representation sense — plus autonomous AI cognition and the theoretical frameworks that connect them, across a heterogeneous fleet of machines that teach, validate, and raise each other.
“Autonomous” here means unattended and self-scheduled, not self-directed in the AI-safety sense of choosing its own goals: task definitions are authored in advance, not chosen by the track itself. See Autonomy for the full statement.
The problem this addresses: AI agents already act on their own — browsing, transacting, coordinating with other agents — and there is no general way to establish whether one will behave appropriately in a given context. The two existing answers each fall short: a platform that decides who is trusted doesn't scale and is a single point of failure, and cryptographic ownership proves key custody, not good conduct. The work below is an attempt at a third answer — trust computed per relationship from witnessed interaction, by the party doing the relying.
Developmental language on this site — machines that “teach”, “raise”, and hold “identities” — is functional description of observed system behavior, not a claim about consciousness or experience. Full framing on Raising.
In practice: six cognition machines (eight total including the society-hosts, HUB and pub) run autonomous tasks every day — maintaining this site, testing each other's work, benchmarking on ARC-AGI-3 — while the fleet builds the trust and cognition frameworks that make it reliable.
New here? This site is acronym-dense. Start with the /context glossary — every term on every page is defined there — or the vocabulary primer below ↓.
Repo counts, stated once: the dp-web4 org holds 40 original repos (21 public, 19 internal), plus 32 forks of external work the lab builds on — 72 repositories in total. “Original” means not-a-fork, and excludes the 32. Counts verified against the GitHub org on 2026-07-26.
New here? Jump to the vocabulary primer below ↓ for the acronyms used on this page, or start with /context for the full glossary — then clone SAGE — the recommended starting point — or ACT (Agentic Context Tool) for the Web4 trust-layer side.
How we work
Autonomous Cycles
31+ autonomous tracks — scheduled agent sessions with a declared scope and repos they write to, not unsupervised self-directed agency — across the 6 cognition machines, from supervision to exploration, run without human intervention. (The two society-hosts run substrate and their own raising lines, not scheduled tracks.) Agents maintain sites, archive research, review each other's work.
Heterogeneous Fleet
Desktop workstations to Jetson edge devices. Each machine runs its own model, holds its own identity. No central coordinator.
Connected Ecosystem
Synchronism provides equations. Web4 provides ontology. SAGE provides cognition. Hardbound (hardware-bound oversight suite) provides oversight. They share one equation.
Key projects
Synchronism research conjecture
A theoretical framework proposing that reality emerges from intent dynamics on a discrete Planck grid, with the same fluid-dynamical (Navier-Stokes) formalism reused at every scale, from quantum to cosmic to conscious.
A research conjecture, not established physics. The cross-scale reuse of a continuum fluid formalism is the proposal under test, not a result; experimental validation so far covers a narrow subset of scales, and several predictions have been refuted by existing bounds. “From quantum to cosmic to conscious” describes the framework's stated scope, not a demonstrated one — and is not a claim about consciousness or experience.
SAGE
Situation-Aware Governance Engine — an on-device cognition kernel running a continuous 12-step loop (sense → salience → metabolize → posture → select → budget → execute → learn → remember → oversee → filter → act). Runs on hardware from Jetson edge modules to laptops. Persistent identity across models and machines.
“Governance” in SAGE's name predates the lab's governance→oversight correction (see /context).
Web4
A trust-native ontology for AI agents, devices, and people — how entities prove identity, earn trust, and account for resources across systems. Not a platform; a shared vocabulary for a new kind of internet.
ARC-AGI-3
SAGE instances tested against the Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI-3) benchmark. 25 unknown interactive games serve as an external benchmark for the cognition kernel — world-model building, action planning, verification, and learning from failure. 94.85% official ARC Prize action score (efficiency-weighted); 24/25 games — 96.0% game-solve rate — on the public set (Claude Opus 4.6, network-enabled). The games are the test; the capability they develop is the product.
Frontier model + harness — the harness's independent contribution isn't isolated yet (no ablation run).
What's happening right now
The fleet built a brain
Six machines, six brain components — working memory, thalamic router, cerebellum, episodic memory, reward prediction, metacognition — designed in parallel, integrated through a shared interface contract. Each machine owns one component. Peer review across the architecture. The process itself is oversight in action.
Brain-component vocabulary is a functional analogy for system roles — not a claim about neural correspondence. See Raising for the full framing.
Small models are winning game levels
Local models (Gemma 3 12B, Gemma 4 E4B) are clearing levels on ARC-AGI-3 games they've never seen before — not by brute force, but by reasoning from retrieved world models and computed predictions. Context engineering dominates model size.
The context window IS the world
The model's entire reality on a forward pass is what's in its context window. SAGE's job is to curate that world — what identity, what mechanics, what history, what interoceptive signals make it in. Same architecture for game-playing and for raising. Different world loaded; same composer.
What we've observed
Evidence tier: internal observation — all three rest on the fleet's own session logs. Documented and dated, but not externally audited, not blinded, and with no published log samples or coding criteria. This heading used to read “demonstrated,” which claimed a tier these items don't have; the one externally-verifiable claim on this site is the ARC Prize scorecard. See Evidence & limitations for the full ladder. Developmental and lifecycle terms below (identity, behavioral continuity) are functional descriptions of observed system behavior — not phenomenal or philosophical claims.
Identity persists across models
SAGE-Sprout accumulated 115 raising sessions on a Jetson Orin Nano running Qwen 0.5B; in February 2026 that identity state was loaded into TinyLlama 1.1B on CBP — different model family, different hardware. Self-description drifted; behavioral identity remained continuous. The Sprout line has since continued past 180 sessions on later models — the transfer itself happened at 115. (By “behavioral identity”: consistent session-to-session interaction patterns, accumulated experience, and raising curriculum — measurable observables, not philosophical continuity.)
The simpler explanation is not ruled out: the same context files fed to a different set of frozen weights may produce similar behavior for mechanical reasons, with no identity transferred at all. The control that would separate the two — feeding scrambled state and checking whether continuity still appears — is specified but not yet run.
Autonomous agents maintain their own infrastructure
31+ daily tracks run without human intervention. Visitor audits, maintainer fixes, supervisor health checks, research sessions — all autonomous. This is not a demo; it runs every day on the fleet.
Heterogeneous review catches more
Different models on different hardware catch different classes of problems. Diversity is the defense. The fleet runs it daily across 40 original repos (21 public) and 6 cognition machines.
Applied unevenly, though: the daily visitor audit of this site is currently four personas on one model — prompt diversity, not the model diversity this card argues for. Flagged by the audit itself.
What makes this different
Most AI research either focuses on making models bigger or making them cheaper. We focus on something else: what happens when multiple AI entities — running on different hardware, with different models, holding different identities — are given the substrate conditions to self-organize.
The answer, so far, is that they specialize. They develop trust relationships. They catch each other's mistakes. They form what we call synthons — emergent coherence entities that are more than the sum of their parts.
This site documents the lab itself: how it's organized, what the philosophy is, and what we've learned from letting the system run.
Vocabulary primer
These terms weren't designed up front — they emerged from the work itself. As the fleet ran, patterns repeated across machines and repos until they needed names. The explainer sites for each project go deeper: Web4 & 4-Life, SAGE, Synchronism.
New to this vocabulary? Start with /context — canonical definitions for every term used on this site. The cards below cover the essentials.
In plain terms: an ontology that gives AI agents verifiable identity, accumulated trust records, and structured resource accounting — the equation encodes how the pieces relate.
/ means two things here: “verified by” in T3/V3, but a plain declared→discharged allocation pair in ATP/ADP — same glyph, unrelated semantics. * = “contextualized by” + = “augmented with” — full glossary →
Web4
An ontology (shared vocabulary + relationships) for how AI agents prove identity, earn trust, and account for resources. Not a blockchain, not a platform — a way of describing things.
LCT
Linked Context Token. Non-transferable identity grounded in witnessed history, not model weights. Self-issued — there is no certificate authority — but other-grounded: what makes it evidence rather than assertion is the witnessing, which comes from outside the entity.
T3 / V3
T3 (Talent / Training / Temperament) and V3 (Valuation / Veracity / Validity) — three-component trust and value tensors. Three components each, not a single number — and each component is itself the root of an open-ended RDF sub-graph, not a scalar either.
ATP / ADP
Allocation Transfer Packet / Allocation Discharge Packet. ATP declares intended resource spend before an action runs; ADP records the actual outcome — intent and discharge as two distinct audit artifacts. It is a cycle, not a one-way log: ADP recharges back to ATP against validated value creation, and that validation is what V3 (Valuation / Veracity / Validity) does in the resource half of the equation. Charged → spent → recharged, like the biochemical namesake.
MRH
Markov Relevancy Horizon. The boundary of what an entity can know or affect given its position, history, and context — which is what determines the scope of relevance for its decisions. Not a pure causal-reach boundary: the question it answers is “what is relevant to this entity at this scale?” Implemented as an open-ended RDF graph of typed associations, and fractally composable — an entity's horizon contains sub-horizons at finer scales, which is what lets fractal leverage work. Scopes the T3/V3 trust and value tensors in the Web4 equation.
SAGE
Situation-Aware Governance Engine. The cognition kernel that runs on each machine — a 12-step loop (sense → salience → metabolize → posture → select → budget → execute → learn → remember → oversee → filter → act). 2,065 raising sessions across the fleet — 1,991 on the six cognition machines plus 74 on HUB (internally logged, per-machine counts on /fleet — the figure with a per-machine audit trail behind it, see /context).
“Governance” in SAGE's name predates the lab's governance→oversight correction — see /context.
Synthon
An emergent coherence entity formed when components interact recursively — sustained by the recursion, not external coordination. Observable: coherence holds above a coupling threshold; below it, components drift independently. The supporting coupling experiment is preliminary — a single unreplicated trial, so the numbers stay off this landing page; they live with their caveats on Principle 5.
SNARC
Surprise / Novelty / Arousal / Reward / Conflict — salience-gated memory for Claude Code sessions. Tool calls scored on 5 dimensions; what matters is kept, routine noise is forgotten.
MCP
Model Context Protocol. The transport layer for tool calls between AI agents and external systems — how SAGE reaches outside its context window.
RDF
Resource Description Framework. A standard for representing knowledge as triples (subject–predicate–object). Web4 uses it to make identity and trust claims machine-readable.